This article appeared in Mind & Language 13, 1998, 421-449.

Archived at Website for the Rutgers University Research Group on Evolution and Higher Cognition.

Theory Theory To The Max(1)

A Critical Notice of: Words, Thoughts and Theories by Alison Gopnik and Andrew N.Meltzoff. Bradford Books: MIT Press. 1997. $30.00

Stephen StichDepartment of Philosophy and Center for Cognitive Science

Rutgers UniversityNew Brunswick, NJ 08901

stich@ruccs.rutgers.edu

and

Shaun NicholsDepartment of Philosophy

College of CharlestonCharleston, SC 29424

nichols@cofc.edu

Contents

  1. Introduction: The Theory that Generates Shocked Incredulity  2. Far Out: Some Dimensions Along Which Theory Theories Differ  3. Some Evolutionary Puzzles  4. The Evidence, The Theory and The Competition  5. Retro Philosophy of Science

1. Introduction: The Theory that Generates Shocked IncredulityOne approach to interdisciplinary theorizing is cautious and deferential, ever aware of the dangersinvolved in making pronouncements about matters on which one is hardly an expert. Typically those whoadopt this strategy propose tentative and elaborately nuanced theories whose complexity echoes thecomplexity of the phenomena to be explained. At the other end of the spectrum are the brash andprovocative interdisciplinary theorists who do not hesitate to defend ambitious, uncomplicated theoriesthat have far reaching implications for fields that are not their own. We doubt there is any way ofknowing in advance which of these approaches will be more successful. But, for us at least, theprovocative approach is usually a lot more fun. Nor is that its only virtue. For when a theory is bold and

straightforward it is usually much easier to see the problems it confronts, and to learn from them.

Those who share our taste for audacious, uncluttered, far reaching theories will find no shortage ofinstructive provocation in Gopnik and Meltzoff's book and in a pair of closely related articles that haverecently appeared in Philosophy of Science.(2)"The theory theory" is the label that G&M adopt for theirview, and "the central idea of this theory is that the processes of cognitive development in children aresimilar to, indeed perhaps even identical with, the processes of cognitive development in scientists."(GM, 3) As Gopnik sees it, "the moral of [the] story is not that children are little scientists but thatscientists are big children. Scientists and children both employ the same particularly powerful andflexible set of cognitive devices." (Ga, p. 486) "[E]veryday cognition, on this view, is simply the theorythat most of us most of the time have arrived at when we get too old and stupid to do more theorizing….We might think of our enterprise as scientists as the further revision of the theory by the fortunate, orpossibly just childish, few who are given leisure to collect evidence and think about it." (G&M, p. 214)Indeed, Gopnik goes on to suggest that "the greatness of individual scientists would literally come fromtheir childishness." (Gb, p. 561)

Suggestions like these have been broached before in the literature by these authors and others (Gopnik,1984 & 1988; Gopnik and Wellman, 1992 & 1994; Wellman, 1985 & 1990; Wellman and Gelman,1992), though, as Gopnik reports, the ideas are often greeted with "shocked incredulity" by "scientists,philosophers and psychologists, particularly those with limited experience of anyone younger than afreshman." (Ga, p. 486) In Words, Thoughts and Theories, G&M attempt to undermine this skepticism byarticulating their position in detail and showing "how it can generate specific predictions, predictions thatare not made by other theories." (G&M, p. 4) They also survey an impressive range of data aboutcognitive development in infancy and childhood, and argue that the theory theory can explain these data.Their book is, by far, the most detailed defense that has yet been offered for the theory theory. Moreoverthe position they advocate is, in a number of ways, an unabashedly extreme version of the theory theory.It's Theory Theory To The Max!

To get a bit clearer on what G&M's version of the theory theory claims, and to see why it is so bracinglysimple and radical, it will be useful to sketch several of the dimensions on which their theory is moreradical than other views to which the "theory theory" label has been applied. That will be our project inthe section to follow. In Section 3, we will briefly consider some of the evolutionary considerations thatG&M offer in support of their theory, and note some further evolutionary considerations that pull in theopposite direction. In Section 4, by far our longest section, we'll provide a more detailed statement ofG&M's theory, sketch some of the facts which, they claim, the theory can explain, and take a critical lookat their arguments aimed at showing that their theory is better than the competition. In the final section,we'll explain why their theory requires them to adopt some extremely controversial views about thenature of scientific change.

2. Far Out: Some Dimensions Along Which Theory TheoriesDifferTo the best of our knowledge, the term "theory theory" was first used by Adam Morton (1980) tocharacterize a cluster of views, in philosophy and psychology, about the ways in which normal adults goabout the business of predicting and explaining the actions of other people and attributing mental states tothem. In all of these activities, according to the sort of theory theory that Morton had in mind, people arerelying on "a fairly extensive commonplace psychological theory, concerned both with dispositional traitssuch as those of character and mood and with the intentions that produce action." (Morton, p. 13) Noteveryone thinks that our "folk psychological" capacity to predict and explain behavior and to attribute

mental states relies on a commonsense theory. Behaviorists don't, of course. Neither do advocates of the"simulation theory," who maintain that in all these activities we gain information about other people byrunning some of our own psychological mechanisms "off-line."(3) But even among those who think thatsome sort of commonsense theory is a crucial part of the explanation of our everyday folk psychologicalskills, there is lots of room for disagreement on what, exactly, counts as a theory. And this is the firstdimension on which G&M stake out an extreme position. For Morton, "not any collection of beliefsforms a theory;"(p. 6) a certain degree of unity is required, and the terms of a theory must be semanticallyinterdependent. But if there is a commonsense psychological theory underlying adult folk psychologicalabilities, it needn't, Morton insists, look much like a scientific theory. "There are," he reminds us,"theories outside science too." (p. 6) Others, including one of the authors of this paper, have taken aneven more permissive view. Our use of commonsense psychological terms, Stich once claimed, "isgoverned by a loose knit network of principles, platitudes and paradigms which constitute a sort of folktheory." (Stich, 1983, p. 1). And in more recent work, where the focus was the plausibility of off-linesimulation theories, we have used the term "theory theory" in a way that (contra Morton) would count justabout any collection of beliefs as a theory. (Stich & Nichols, 1992 & 1995) But G&M have no truck withthis insipid inclusiveness. They offer a detailed account of the structural, functional and dynamic featureswhich, they claim, characterize most scientific theories (G&M, pp. 32-41), and they go on to claim thatthe cognitive structures subserving folk psychological skills in both children and adults have all of thesefeatures. More generally, when G&M claim that a given body of knowledge in any domain is theoretical,they mean that it has all of the characteristics they attribute to scientific theories.(4)

A second way in which G&M embrace a stronger position than many others who would call themselves"theory theorists" turns on the range of commonsense capacities which are said to be subserved by atheory. For Morton, as we've seen, the theory theory is a hypothesis about commonsense psychologicalcapacities. But G&M extend the hypothesis to cover three large domains of commonsense knowledge.The first of these, which they call the "theory of object appearances," includes knowledge about themovement of objects, the properties of objects, the spatial relations between stationary objects and theperceptual relations between observers and objects. (G&M, p. 78) On G&M's account, "[t]he theory ofobject appearances is an interesting bridge between 'folk physics' and 'folk psychology.' It involves ideasabout objects and the relations among them, but just as crucially, it involves ideas about the way that weand others perceive objects." (G&M, p. 78) The second domain in which G&M maintain thatcommonsense knowledge is theoretical is what they call "the theory of action." This includes much ofwhat is usually classified as commonsense psychological information, including information about howbeliefs and desires arise and change and how they interact to produce actions. It also includes informationabout the ways in which actions have effects on both the physical world and on people. The third domainthat G&M discuss is "the theory of kinds," which specifies the principles according to which objects areto be grouped together into kinds and the sorts of inferences to be drawn from the fact that objects are ofthe same kind. Though these are the three domains that G&M treat in detail, they do not preclude thepossibility that knowledge in other domains is also theoretical.(5) On the other hand, they make it clearthat they do not think all commonsense knowledge is theoretical. Rather, they expect that scripts,narratives, empirical generalizations and other forms of information packaging will all have a role to playin explaining some aspects of commonsense knowledge.

Thus far we've been considering what G&M's version of the theory theory has to say about the knowledgestructures that underlie a range of normal adult capacities. But since both Gopnik and Meltzoff aredevelopmental psychologists it is not surprising that the boldest and most radical aspect of their theory isthe position they take on developmental questions -- questions about how these adult knowledgestructures emerge. Here there are two issues to consider. First, when exactly does our knowledge in therelevant domains start being theoretical? As we trace a person's knowledge in these domains back fromadulthood to adolescence through childhood to infancy, when do the first theories appear? The

astonishing answer that G&M propose (with obvious delight) is that the child's knowledge in the threedomains they consider is always theoretical; right from the start it has all the features that are fundamentalto scientific theories.

At this point some of the psychologists, scientists and philosophers who were crying out …may well be doing so again; "Surely, you cannot think it is theories all the way down!" Well,yes, actually I do think it is theories all the way down…. Infants seem to have innateknowledge … and this knowledge is theory-like…. (When we say that this knowledge isinnate we do not mean this in the philosophical sense, which is that neither the philosopher inquestion nor any of the guys down the hall could think of a way to learn it. We mean that ithas been demonstrated in 42-minute-old infants.) (Ga, p. 510)

Most previously published applications of the theory theory have involved older children, including somewho are actively learning scientific concepts. But these studies have no direct bearing on what G&M dub"initial state nativism" -- the claim that infants are born knowing theories. One of the main aims of theirbook "is to apply the theory theory to explain what we know about infancy and very early childhood,"(G&M, p. 4) and thus much of the evidence they recount deals with cognitive development in the first 36months of life.

A second developmental question focuses on the mechanisms that are responsible for the theoreticalrevisions and replacements that (they maintain) occur during the course of cognitive development inchildhood. What are these mechanisms? How exactly do they work? Here G&M concede that they do nothave a complete story to tell. (G&M, p. 218 ff) What they do claim, as we've already noted, is that themechanisms responsible for theory change in infants and children are exactly the same as the mechanismsthat are responsible for theory change in science. According to G&M, human beings have only one set ofmental mechanisms for theory revision. These mechanisms play a major role in the cognitive changes thatmark everyone's infancy and childhood. In the privileged few who have the leisure to pursue science asadults, the mechanisms continue to be used.

3. Some Evolutionary PuzzlesOne virtue that G&M claim for their version of the theory theory is that it solves "an interestingevolutionary puzzle." Everyone agrees that humans (at least some humans) have the capacity to doscience. And few would challenge the claim that in doing science people use a flexible and powerful setof cognitive abilities. But, G&M ask, "Where did the particularly powerful and flexible devices of sciencecome from? After all, we have only been doing science in an organized way for the last 500 years or so;presumably they didn't evolve so that we could do that." (G&M, p. 18 & Ga, p. 489) The answer theysuggest is that

many of these cognitive devices are involved in the staggering amount of learning that goeson in infancy and childhood. Indeed, we might tell the evolutionary story that these devicesevolved to allow human children, in particular, to learn. (G&M, p. 18 & Ga, p. 489)

From an evolutionary point of view, three of the most distinctive features of human beingsare the plasticity of their behavior, their ability to adapt to an extremely wide variety ofenvironments and their long, protected immaturity. Equipping human children withparticularly powerful and flexible cognitive devices, devices that are good at constructingaccurate representations of new and unexpected worlds, might be an important part of thisevolutionary strategy. We might indeed think of childhood as a period when many of therequirements for survival are suspended, so that children can concentrate on acquiring a

veridical picture of the particular physical and social world in which they find themselves.Once they know where they are, as it were, they can figure out what to do. On this view wemight think of infancy as a sort of extended stay in a Center for Advanced Studies, with evenbetter food delivery systems….

[T]hese powerful theory formation abilities continue to allow all of us at some times, andsome of us, namely professional scientists, much of the time, to continue to discover moreand more stuff about the world around us. On this view science is a kind of spandrel, anepiphenomenon of childhood. (Ga, p. 490)

This proposed solution to the evolutionary puzzle gives rise to two further puzzles, one of which Gopnikacknowledges and addresses with considerable ingenuity. The other goes unnoticed.

The puzzle that Gopnik acknowledges is what Giere (1996, p. 539) calls "the 1492 problem." "Science aswe know it," Giere notes, "did not exist in 1492." But if G&M are right, then the cognitive devices thatgive rise to science have been part of our heritage since the Pleistocene. So why have humans only beendoing science for the last 500 years? The answer, according to Gopnik, is mostly a matter of theavailability of relevant evidence. "My guess is that children, as well as ordinary adults, do not …systematically search for evidence that falsifies their hypotheses, though … they do revise their theorieswhen a sufficient amount of falsifying evidence is presented to them. In a very evidentially rich situation,the sort of situation in which children find themselves, there is no point in such a search; falsifyingevidence will batter you over the head soon enough." (Gb, p. 554) Now what happened about 500 yearsago, Gopnik maintains, is that as a result of various historical and social factors a few thinkers foundthemselves confronted with unprecedented amounts of new evidence relevant to several venerablequestions like: Why do the stars move as they do? New technology was one reason for the availability ofnew evidence; telescopes were invented. Other technological and social changes greatly facilitatedcommunication allowing "a mathematician in Italy to know what an astronomer has seen in Denmark."(Gb, p. 554) Greater leisure (at least for a few) was yet another factor, and so too was the emergence ofthe experimental method which motivated those who adopted it to systematically seek out new andpotentially falsifying evidence. All of this, and perhaps other factors as well, created an environment inwhich the theory revision mechanisms that natural selection had designed to enable children to cope with"new and unexpected worlds" might begin functioning actively in adulthood, long past the stage in life inwhich they would have made their principal contribution to fitness in the environment in which ourancestors evolved.

There is, we think, some plausibility to this story. It meshes nicely with recent accounts that stress theimportance of environmental instability in hominid evolution. (See, for example, Potts, 1996.) If, as theevidence suggests, climactic changes became much more frequent during the last 2.8 million years,hominids that were good at learning to deal with radically changed environmental conditions might wellhave had a considerable selective advantage. But there is also a significant, and unacknowledged, tensionbetween this account and one of the more interesting and extreme features of G&M's view. Recall thataccording to G&M "it's theories all the way down" and that even the very early development of the child'sknowledge about object appearances, actions and kinds is driven by a process of theory revision -- thevery same process whose evolutionary function is to enable older children (and scientists) to acquire "averidical picture of the particular physical and social world in which they find themselves" and thus "toadapt to an extremely wide variety of environments." (Ga, p. 490) Now what is puzzling about this is thatjust about all the knowledge that children are acquiring in the early stages of this process - knowledgeabout the continuous movement of objects, for example, or about the spatial relations that must obtainbetween an agent and an object if the agent is able to see the object, or about the way in which an agent'sdesires lead to actions - deals with aspects of the environment which are not in the least variable. Thesefacts have been fixed and unchanging since long before the emergence of hominids or primates. Indeed,

some of these facts (particularly those about the movements and interactions of middle sized physicalobjects) have presumably obtained in the environment of every organism that has ever existed on Earth.Moreover, it is also the case that knowledge of these facts would be adaptive even to creatures which didnot have to cope with a highly variable environment and which (because of this, or for whatever reason)never evolved anything like the human capacity to develop veridical theories about a wide variety ofenvironments. Thus it is hardly surprising that, as Carey and Spelke (1996) note, infants and primatesreason similarly about objects, or that young children and other mammals construct similarrepresentations of space. There is even some evidence suggesting that 2-day-old chicks perceive thecomplete shapes of partly hidden objects and track the location of fully hidden objects in much the sameway that human infants do.

Obviously, the mechanism responsible for the young chicken's acquisition of the ability to track thelocation of hidden objects must be very different from the powerful and flexible cognitive devices thatenable humans to construct accurate representations of new and unexpected worlds. Moreover, it isoverwhelmingly likely that our own primate and pre-primate ancestors had the knowledge necessary todeal with the movements and trajectories of physical objects long before the appearance of the sort oftheory revision mechanisms posited by G&M. So there must have been some mechanisms in place in ourprimate forebears which explained their possession of this knowledge, and those mechanisms could nothave been the flexible and powerful theory revision devices. Yet if G&M are right, the theory revisiondevices are the mechanisms responsible for the acquisition of much the same knowledge in humaninfants and young children. But it is puzzling, to put it mildly, how this change might have come about.Why would natural selection abandon a perfectly good and extremely reliable system for generatingknowledge about certain utterly stable and unchanging parts of the environment, and assign the task to anew, more fallible system whose main adaptive virtue is that it can gain knowledge about highly variableparts of the environment?(6) If the older, inflexible system ain't broke, why on earth would naturalselection fix it? Wouldn't it make more sense for natural selection to retain the older system for the earlyacquisition of knowledge about stable features of the world, and arrange for the theory revision device tokick in later on in development when children must figure out which of the many possible environmentsthey happen to have ended up in? A two tiered system of this sort could still tell the story G&M tell aboutscience being a spandrel, though it would be an epiphenomenon of a somewhat later stage of childhood.What this two tiered hypothesis rejects is G&M's radical contention that it's theories (and theory revision)all the way down.

These evolutionary considerations hardly constitute a knockdown argument against G&M's version of thetheory theory. There are lots of stories to be told about why natural selection might dump the system thatyields knowledge about spatial relations and the motions of physical objects in non-human primates andreplace it with a theory revision system of the sort that G&M propose. Perhaps it's easier to integrateprimitive and more sophisticated knowledge if the two are generated by the same system; perhapsmaintenance costs are lower if there is only one system, rather than two. This game is easy to play. Theconclusion we would draw is that while G&M's theory is compatible with lots of the facts that needexplaining, the fit is far from perfect, and there are alternative theories that do at least as good a job atexplaining the facts. This conclusion will become a leitmotif of the section to follow, where our centraltheme will be that G&M do not make a convincing case for the claim that their theory does a better jobthan the competition at explaining the evidence.

4. The Evidence, The Theory and The CompetitionThe central claim of G&M's theory is that the processes of cognitive development in children, includingvery young children, are similar or identical to the processes of cognitive development in science. Tomake a case for this claim, G&M adopt the following strategy. First, they give an account of the

processes of cognitive development in science. Next, they recount a great deal of experimental andanecdotal evidence about cognitive development in children, and argue that all of these facts arecompatible with their theory -- if the theory were true it would explain the experimental and anecdotalfindings. But, as G&M are aware, this by itself would only be enough to show that their theory is aserious contender. In order to argue that their theory offers the best explanation of the available facts theymust, and do, argue that there are facts which competing theories cannot comfortably explain. It's ourcontention that their arguments for this last claim are singularly unpersuasive. To explain why we findtheir arguments unconvincing, we will proceed as follows. First, we'll sketch the account of scientifictheories and their development that G&M adopt. Next, we'll offer a few illustrations of the sorts ofevidence about young children that G&M claim can be explained by their theory. We'll then consider insome detail the arguments that G&M offer for the claim that their theory does a better job at explainingthe facts than competing theories.

4.1. An Account of Scientific Theories & Their Development.

There is, as G&M note, considerable disagreement among philosophers, historians and sociologists ofscience about just what scientific theories are and how they change. Confronted with this controversy,G&M propose to take "the modest and emollient route of focusing on those features of theories that aremost generally accepted across many different conceptions of science," since they want to be "asmainstream and middle-of-the-road as possible."(G&M, p. 33) This is, we think a fair characterization ofmuch of what they say about science, though in Section 5 we'll note that some of the claims their theoryrequires them to make about science are far from mainstream, and far from plausible.

In setting out their account of theories, G&M focus on three sorts of features - structural, functional anddynamic - which they claim are distinctive of scientific theories. The four "static structural" features theymention are:

(1) Abstractness: "[T]heoretical constructs are typically phrased in a vocabulary that isdifferent from the vocabulary of the evidence that supports the theory." (G&M, p. 34)

(2) Coherence: "The entities postulated by a theory are closely, 'lawfully,' interrelated withone another." (G&M, p. 35)

(3) Causality: "[I]n theories we appeal to some underlying causal structure that we think isresponsible for the superficial regularities in the data…. The intratheoretic relations, the laws,are typically interpreted in a causal way …. [and] the theoretical entities are seen to becausally responsible for the evidence. (G&M. p. 35)

(4) Ontological commitment: "[T]heories make ontological commitments and supportcounterfactuals…. [Thus] a test of theoreticity … is the nature of our surprise at violations ofthe theory. If we are committed to the theory, such violations strike us not only as surprisingbut as being impossible and unbelievable in an important and strong way. This differentiatestheories from other types of knowledge." (G&M, pp. 35-6)

The functional features, the things that theories do (or that people do with them) are:

(5) Prediction: "A theory, in contrast to a mere empirical generalization, makes predictionsabout a wide variety of evidence, including evidence that played no role in the theory'sconstruction." (G&M, p. 36)

(6) Interpretation: "[T]heories strongly influence which pieces of evidence we considersalient or important." (G&M, p. 37)

(7) Explanation: "The coherence and abstractness of theories and their causal attributions andontological commitments together give them an explanatory force lacking in mere typologiesof, or generalizations about, the data." (G&M, p. 38) And why do people seek explanatorytheories? Evolution built us with the motivation to explain because that motivation leads usto build theories, and good theories enable us to deal more effectively with our environment."From an evolutionary point of view we might suggest that explanation is to cognition asorgasm (or at least male orgasm) is to reproduction." (G&M, p. 38)

Though all of these features play a role at one point or another in G&M's argument for the theory theory,much of the weight of their argument rests on the dynamic features of theories:

(8) Defeasibility: "[T]he most important thing about theories is what philosophers call theirdefeasibility. Theories may turn out to be inconsistent with the evidence, and because of thistheories change." (G&M, p. 39)

(9) The crucial role of counter-evidence: "Theories change as a result of a number ofdifferent epistemological processes. One particularly critical factor is the accumulation ofcounterevidence to the theory." (G&M, p. 39)

(10) Characteristic intermediate processes: "There are characteristic intermediate processesinvolved in the transition from one theory to another." (G&M, p. 39)

(i) Denial: "The initial reaction of a theory to counterevidence may be a kind ofdenial. The interpretive mechanism of the theory may treat the counterevidenceas noise, mess, not worth attending to." (G&M, p. 39)

(ii) Auxiliary hypotheses: "At a slightly later stage the theory may develop adhoc auxiliary hypotheses designed to account superficially for thecounterevidence…. But such auxiliary hypotheses often appear, over time, toundermine the theory's coherence…. The theory gets ugly and messy instead ofbeing beautiful and simple. The preference for beautiful theories over ugly ones(usually phrased, less poetically, in terms of simplicity criteria) plays anadditional major role in theory change." (G&M, p. 39)

(iii) Appearance of an alternative: "The next step requires an alternative modelto the original theory…. The ability to fix on this alternative is the mysteriouslogic of discovery." (G&M, p. 40)

(iv) Intense experimentation & observation: "A final important dynamic featureof theory formation is a period of intense experimentation and/or observation….The role that experimentation and observation play in theory change is stillmysterious, but that it plays a role seems plain." (G&M, p. 40)

4.2. The Theory Meets the Evidence

The version of the theory theory that G&M want to defend claims that children, including very youngchildren, have and use theories with the structural and functional features sketched in (1) - (7), and that inthe course of development these theories are replaced with better theories by a process that fits the patternelaborated in (8) - (10). This is not the place to attempt a detailed summary of experimental findings thatG&M recount, though much of this evidence is fascinating and surprising, and the tour they offer is wellworth the price of the book even if you are no more convinced by their theory than we are. But whilethere can be little doubt that the results they discuss are intriguing and important, it is often a bit of a

stretch to see why they think the evidence supports their theory.

Consider, for example, Meltzoff and Moore's (1983, 1989) amazing finding that infants as young as 42minutes old are able to imitate facial gestures such as mouth opening, tongue protrusion and lipprotrusion. Since infants cannot see their own faces, G&M argue that this capacity requires an innate,cross-modal mapping. "There is," they maintain, "an abstract representation, a kind of body scheme, thatallows an innate mapping from certain kinds of behavioral observations of others to certain kinds ofperceptions of our own internal states. In particular, we innately map the visually perceived motions ofothers onto our own kinesthetic sensations." (G&M, p. 129) All of this is plausible enough. But whatbearing does it have on the claim that "infants have innate knowledge … and this knowledge is theory-like." (Ga, p. p. 510) Well perhaps, as Gopnik suggests, the abstractness of the representation thatsubserves the cross-modal mapping shows that the child's innate knowledge has the first of the tenfeatures on the list set out in 4.1. Gopnik also maintains that we can view these infants as "drawing atleast a primitive kind of inference and prediction" (Ga, p. p. 510) on the basis of this knowledge, thusgetting us feature (5). We're inclined to think that this requires a rather generous interpretation of thefacts. For what exactly does Gopnik think the infant is predicting? The most obvious suggestion is that heis predicting that kinesthetic sensations of a certain sort will produce behavior similar to the behavior hehas just observed. And we find it a bit hard to take this proposal seriously. But even if we concede thepoint on feature (5), what about the other 8 features? What reason do we have to suppose that the infanttakes the entities posited by his theory (whatever exactly these are) to be "closely, lawfully, related withone another," or that the infant is interpreting anything in a causal way, or that his knowledge supportscounterfactuals, or that what he knows has an explanatory force? The answer, as best we can see, is thatwe have no reason at all to suppose any of this. Of course, this hardly shows that G&M's theory ismistaken. Perhaps the infant is making lots of causal interpretations, contemplating counterfactuals andgenerating orgiastic explanations one after another, and we just haven't yet figured out how todemonstrate this experimentally. Or perhaps, though the infant has a theory which would support all ofthese cognitive activities, he's just not much inclined to engage in any of them at this very tender age. Thepoint here is that while G&M's theory may be broadly compatible with the evidence they cite, the theoryclaims vastly more than the evidence gives us any reason to believe. Moreover, the evidence is equallycompatible with a wide range of alternative hypothesis which do not attribute to 42-minute-old infants aknowledge of theories with all the fundamental properties of scientific theories.

For a second example let us consider G&M's intriguing proposal about the much studied A-not-B errorthat infants make between the ages of 9 and 12 months. During the first 9 months of life an infant'sknowledge about the motions of objects evolves in complex ways. By 6 months gaze trackingexperiments show that infants can project the visible trajectory of an object even when it disappearsbehind a screen. But they seem quite untroubled if an object smoothly moves from left to right,disappears behind the left edge of one screen, and reappears at the right edge of a second screen withoutever appearing in the gap between the two screens. Also, at this age, they fail to search for an object, evena highly desirable one like a favorite toy, when it disappears behind a screen or a cloth. By 9 months thesituation has changed considerably. While their gaze will still track the trajectory of an object as it passesbehind a screen, if the object fails to appear at the gap between one screen and another their tracking willbe disrupted and they will look back to the first screen. Also, at 9 months, they have little trouble findingan object hidden under a single cloth. If an object hidden under a cloth disappears, "they even shake thecloth as if they expect the object to appear there. All this suggests that these infants genuinely postulatethat the object will be behind the occluder where it disappeared." (G&M, p. 95) Oddly, however, a nine-month-old will often make what is known as the A-not-B error: If the infant observes an object beinghidden and recovered several times under cloth A and then observes the object being hidden under adifferent cloth, B, the infant searches persistently under cloth A rather than under cloth B.

The explanation that G&M propose for these and other facts is as follows. At 9 months, children have "a

rich and abstract conception of objects." (G&M, p. 98) They "believe that the object is at the place ortrajectory at which it disappeared and that it is behind the occluder and that the fact that it is behind theoccluder is what makes the object invisible." (G&M, p. 95) However, there are many common situationswhich this theory simply does not handle. These include situations in which the object is "invisiblydisplaced in the sense that it changes its trajectory and traces an invisible path of movement that cannotbe extrapolated from its visible path." (G&M, p. 97) This is what happens, for example, when a childthrows her teddy bear out of her crib (and out of sight) and, with some help from Mom, it later ends up inthe toy box. "What can the baby do in these circumstances? There is a rule that can handle the teddy bearcase and many others where invisibly displaced objects reappear at places that were not part of theiroriginal trajectory. Moreover, the rule doesn't require that you have a theory of invisible displacements.Many objects, especially in the child's world, have habitual locations. The rule is, 'The object will bewhere it appeared before.'" (G&M, p. 99) This rule "will receive a great deal of empirical confirmation.We might think of it as an example of a purely empirical generalization…. But from the point of view ofthe adult theory, and arguably of any coherent theory, both this generalization and the theoreticalprediction cannot simultaneously be true; the object can't be under both cloth A [where it appearedbefore] and under cloth B [where it disappeared]…. In fact, the evidence suggests that infants may alsodevelop other often contradictory empirical generalizations in the 9-to-15-month period." (G&M, pp. 99-100)

For G&M, the payoff in all of this is the link with (10, ii) on the list of the features of scientific theories:"This situation is analogous to similar situations in science where, without a theory to resolve them,contradictory generalizations may proliferate. The 9-month-old is like a scientist who attempts to save thetheory by adding ad hoc auxilliary hypotheses. The central theory of object movements leads to apparentanomalies when objects are invisibly displaced. The ad hoc rule 'It will be where it appeared before' isinvoked to deal with these anomalies." (G&M, p. 100) By about 12 months, children stop making A-not-B errors, and this leads G&M to claim that they have abandoned the ad hoc auxilliary hypothesis. "Whydo they abandon it? The fact that the rule is sometimes disconfirmed, as in the A-not-B situation,probably plays a part. We suspect, however, that the lack of consistency between this rule and thecentrally developed theory of objects that leads to the rule 'It will be where it disappeared' plays an evenmore important role. To return to our earlier account, the child is analogous to a scientist who is disturbedto discover that her ad hoc auxilliary hypotheses lead to contradictory predictions."(G&M, p. 100)

As we see it, the appropriate conclusion to draw from this example is quite similar to the conclusion wedrew from the previous one: the evidence that G&M cite is indeed broadly compatible with their strongversion of the theory theory. However here as before one has to be rather generous in interpreting thetheory to get it to square with the facts. One of these facts is that, as G&M note (G&M, p. 96), undercertain circumstances 9-month-olds will make A-not-B errors even when the object at B is still clearlyvisible. Presumably the explanation that they would offer for these cases is that the child's reliance on thead hoc auxilliary hypothesis is so strong that it overrules not only theoretical predictions about invisibleobjects but also direct observations of visible ones. It's possible, we suppose, though surely it's a bit of astretch. Moreover, as both G&M and Carey and Spelke (1996, pp. 521-2) note, there are otherexplanations in the literature that trace A-not-B errors to maturational changes in the brain structuressubserving means/end planning and the inhibition of competing responses. Carey and Spelke also notethat "Diamond … has shown that the developmental changes involving the A/not B errors of infants of 7months and beyond, are mirrored, in parametric detail, by identical changes in 2- to 4-month old rhesusmonkey infants." (Carey and Spelke, 1996, p. 521) We doubt that G&M would explain this finding byattributing ad hoc hypotheses to rhesus monkeys, so they would have to argue that, despite the strongparallels, the processes underlying A-not-B errors in humans is radically different from the processes atwork in monkeys.

We've considered only two examples drawn from the impressive body of evidence that G&M assemble.

However, what we've said about these two cases can, we think, be said about much of the rest of theirevidence: with a bit of squinting it can all be seen as broadly compatible with their theory. And the factthat their theory is compatible with such a large and varied body of developmental evidence is, we think,more than enough to establish it as a real contender. Despite the "shocked incredulity" with which thetheory is often greeted, G&M have made an impressive case that their strong, theories-all-the-way-downversion of the theory theory must be taken very seriously indeed. But of course G&M would hardly besatisfied with this. They claim that their theory is not only a serious contender but the leading contender.To defend this much stronger claim they have to argue that their theory is better than the availablealternatives. And argue they do. As we see it, however, these arguments are by far the weakest parts oftheir case.

4.3. The Theory Meets the Competition

As G&M view it, the principal competitors to the theory theory fall into two broad categories. One ofthese, which they call "empirical generalizations" includes "scripts, narratives, connectionist nets, andother cognitive structures quite closely related to immediate experience." (G&M, p. 50) Since theyrecognize that both the potential of connectionist models and the exact nature of the claims made for themare far from clear (G&M, p. 218), G&M quite sensibly do not mount a detailed critique of such models.They do, however, argue that scripts, narratives and similar structures are simply insufficiently abstract toexplain the wide range of predictions, including predictions about entirely novel cases, that youngchildren can make about objects and actions. We think the case they make is quite convincing.

For the other category of competing theories, G&M use the label "modules." As they characterize them,modularity theories claim that "representations of the world are not constructed from evidence in thecourse of development. Instead, representations are produced by innate structures, modules, or constraintsthat have been constructed in the course of evolution. These structures may need to be triggered, but oncethey are triggered, they create mandatory representations of input." (G&M, p. 50) "The classic examplesof modules are the specialized representations and rules of the visual and syntactic systems." (G&M, p.51) Other examples that G&M discuss include the "core knowledge" theory advocated by Spelke, Careyand others, Leslie's account of the "theory of mind mechanism," and Cosmides and Tooby's theory aboutthe cognitive mechanism that subserves reasoning about permission and obligation. G&M do not claimthat modules play no role in cognition. Quite the opposite. They advocate "a kind of developmentalpluralism: there are many quite different mechanisms underlying cognitive development." (G&M, p. 49)However, they maintain that "theory formation rather than these other mechanisms accounts for theparticular cognitive and semantic phenomena" that they describe in their discussion of the theory ofobject appearances, the theory of action and the theory of kinds. (G&M, p. 50) They also make it clearthat what divides them from modularity theorists is not a dispute over nativism, for "while modules areinnate, not all innate structures are modules," (G&M, p. 51) and on their version of the theory theory,infants are born with innate theories. Unlike modules, these innate theories are "defeasible; any part ofthem could be, and indeed will be, altered by new evidence." (G&M, p. 51)

4.3.1. An argument based on the static properties of theories

How are we to tell whether the mental representations that underlie the child's developing skills in dealingwith objects, agents and kinds are the product of innate modules or of processes of theory revision of thesort that G&M champion? There are, it seems, two sorts of evidence that we might consider. The first isevidence that bears on the synchronic or static features of these representations. Since G&M claim thatit's theories all the way down, we should expect that the representations subserving the child's skills withobjects, agents and kinds always manifest the first seven features on the list set out in 4.1, and as we havenoted, G&M do cite some evidence that this is the case even very early on. However, when the goal is toshow not merely that the theory theory is compatible with the evidence, but that it is better than the

module alternative, evidence bearing on the static properties of the child's mental representations simplyis not relevant. For while a modularity theorist might claim that the representations underlying one ofthese skills lacks some of the features (1)-(7), a modularity theorist might also claim that some innatemodules exploit representations that are identical with theories in all of their static properties. So whileevidence bearing on the static features of the representations underlying skills with objects, agents andkinds could in principle falsify G&M's theory, such evidence could not falsify the modular alternative.Most of the time G&M see this point very clearly. For example, in discussing "what kinds of evidencecould differentiate between a modularity theory and a theory theory that includes innate theories," theynote that

it may be difficult, if not impossible, to distinguish these views by looking at a single staticrepresentational system. At least some of the structural and functional features of theories -their abstractness, coherence, and predictive interpretive force - can also be found inmodules. (G&M, p. 52; for similar passages, see pp. 50 & 90)

On a number of occasions, however, G&M seem to forget the point and argue that static properties cancount against modular theories. Here is an example:

One test of whether a particular belief was the result of a module, an empiricalgeneralization, or a theory might be to think about how we would react to an event thatviolated that belief. If our knowledge was really modular in a strict sense, we should not beable to represent the event at all, as our perceptual system can't represent the reality in aperceptual illusion and children can't represent the syntax of a pidgin language. (G&M, p. 79)

They then go on to argue (G&M, pp. 80-81) that the fact that we can "override" the "perceptualrepresentations" generated by our perceptual system in the case of illusions poses a problem for Spelke'smodular "core knowledge" account of our knowledge of objects. What's going on here? Since theirargument at this point is not up to their usual standards of clarity, it's hard to be sure. But here's our guess.First of all, note the curious phrase "modular in the strict sense." What exactly does it mean? ThoughG&M never define it, the context suggests the following account: If a belief or "perceptualrepresentation" is the product of a body of knowledge which is "modular in the strict sense" then thatrepresentation cannot be overridden by "further evidence." To override a representation is to come tobelieve something other than what the representation claims to be the case. Thus if a strictly modularsystem generates a claim about some state of affairs, then we can't believe any alternative claim, indeed,perhaps we can't even conceive of an alternative. Now, with all this in place, they are in a position toargue that if some perceptual representation or belief can be overridden, then that representation cannotbe the product of a system of knowledge that is "modular in a strict sense." So if "in ordinary life there aremany cases where our predictions about objects and disappearances would plainly have to override anyinnate core principles," (81) that spells trouble for the view that our knowledge about the behavior ofobjects includes a system of core principles that is modular in the strict sense. Are there such cases inordinary life? Of course there are:

To use a simple and ubiquitous example, when we eat things, we do not assume that they willappear at the location where they disappeared. Instead, we assume they will be transformedor destroyed. (81)

And thus it is that commonsense knowledge about the ultimate fate of what we eat refutes Spelke's theoryof core knowledge.

Readers who have even the smallest sympathy with the principle of charity in interpretation are bound tofeel rather uncomfortable with this argument against Spelke. Could it really be the case that one of the

world's leading developmental psychologists has proposed a theory that can be refuted by the fact thatmost people know that what they eat is transformed into something else? The answer, not surprisingly, isno. Moreover, it is pretty clear where G&M's argument goes wrong. What they establish is that if ourknowledge of object movements includes a core knowledge system of the sort that Spelke posits, then thiscore knowledge is not modular in the strict sense. Once we see what this means, however, it is clear thatG&M are attacking a straw man. A body of knowledge is modular "in the strict sense" only if we have tobelieve all the representations of the world that this body of knowledge generates. And Spelke makes nosuch claim about the representations of the world generated by core knowledge. She makes it clear thatthe core knowledge systems she posits are often response and task specific. Typically, the job of therepresentations they generate is not to determine what we believe but rather to guide some very specificactivity. Thus, for example, Spelke maintains that different systems of core knowledge may be guidingvisual tracking and predictive reaching, and under certain circumstances the predictions that thesesystems generate will be incompatible. (Spelke, 1994, p. 441; Carey and Spelke, 1996, p. 519) Obviously,we don't end up believing both predictions, and in some circumstances we won't end up believing eitherone. In determining what we believe, all things considered, the pronouncements of core knowledgesystems are eminently overrideable, and Spelke herself has been quite clear about this.(7)

4.3.2. Three arguments based on the dynamic properties of theories

Let us turn now to the second sort of evidence that might be invoked to determine whether the theorytheory or a modularity theory does a better job at explaining the child's capacity in dealing with objects,agents and kinds. This, it will be recalled, is evidence about the dynamic properties predicted by the twosorts of theories. On the view that G&M urge (most of the time) it is this evidence that will be crucial.

[Modular representations] will not have the dynamic features of theories. In particular, theywill be indefeasible; they will not be changed or revised in response to evidence. (G&M, p.50)

The crucial evidence differentiating the two views lies in the dynamic properties of modulesand theories, in how they develop. (G&M, p. 52)

What kinds of facts about the ways in which the representations or knowledge structures underlyingchildren's abilities change over time do G&M think would count in favor of the theory theory and againstmodularity theories? As we read them they make three quite different suggestions.

The first suggestion: The child's representations change and become more accurate.

The first suggestion is that the very fact that these representations do change, and that they generally getbetter or more accurate over time, counts in favor of the theory theory and against the modularityaccounts.

Modular representations do not lead to predictions through some set of inductive anddeductive generalizations or through a process of theory testing, confirmation anddisconfirmation. They lead to predictions because they are specifically designed by evolutionto do so.

A consequence of this is that modularity theories are, in an important sense,antidevelopmental. Apparent changes in representation occurring over time, on these views,can be accounted for only by processes outside the representational system itself. Onepossibility is that they reflect the maturation of another innate structure, a later modulecoming on line....

Performance deficits are also often invoked to deal with cases in which the child apparentlyhas incorrect representations at one point, which are replaced by other more accuraterepresentations later on. Such sequences are predicted by the theory theory. They areanomalous, however, for modularity theories. It is easy to see why evolution might havedesigned a representational system that was inaccurate in some respects. It is much moredifficult, however, to see why evolution would have designed a sequence of incorrectmodules, each maturing only to be replaced by another. (G&M, pp. 54-5; emphasis added)

Since this is not a theme that G&M pursue elsewhere in their book, we suspect that they don't really wantto place much weight on the argument, and that's all to the good since the argument clearly won't bearmuch. It is, after all, a well established principle of developmental biology that ontogeny oftenrecapitulates phylogeny.(8) Early in their development, primate fetuses have structures that resemble gills;as development proceeds, these structures disappear and are replaced by embryonic lungs. In this andmany other standard examples, it appears that in the course of development organisms pass throughstages that resemble the adult stage of their evolutionary forebears. Since this pattern is common indevelopment, the modularity theorist is surely not going to be in the least embarrassed to propose thathuman children develop a sequence of incorrect modular representational systems each of which isreplaced by a more accurate system as the child matures. These inaccurate systems may simply be thedevelopmental echoes of adult systems in the organisms from which we evolved.

The second suggestion: If given the opportunity, children would learn about radically different worlds

A second sort of fact which G&M claim would count in favor of the theory theory and against modularitytheories would be a demonstration that children raised in a physical or psychological environment that isvery different from ours end up with a correct theory about that environment, just as children raised in ourenvironment end up with a correct theory about our environment.

There is, in principle, a simple experiment that could always discriminate modularity theoryand theory theory. Place some children in a universe that is radically different from our own,keep them healthy and sane for a reasonably long period of time, and see what they come upwith. If they come up with representations that are an accurate account of our universe,modularity is right. If they come up with representations that are an accurate account of theiruniverse, the theory theory is right. Unfortunately, given the constraints of the federal budget,not to mention the constraints of conscience, this experiment is impossible. (G&M, p. 53)

This sort of "in principle" crucial experiment is a recurrent theme in G&M's defense of the theory theory.On three occasions (pp. 81-2, 127-8 & 165) they elaborate on the proposal by describing aspects of thewonderfully weird worlds portrayed in the Star Trek series, and predicting that human children raised inthose environments would develop accurate theories about them - theories very different from the onesthat human children develop in our world. The Star Trek thought experiments are delightfullyentertaining, so much so, indeed, that they led one of us to go out and rent the most recent Star Trekmovie. (And much to his surprise, it is also delightfully entertaining!) However, it is our contention thatin proposing these experiments as definitive (albeit practically impossible) ways to determine whether thetheory theory or the modularity theory is correct, G&M are deeply confused.

A first problem with the proposal is the suggestion, made quite explicitly in the passage quoted above,that if modularity theory is correct, then all children who are sane and healthy will come up with the sametheory (one that happens to be true of the world in which we live) no matter what world they inhabit. Thismight indeed be the case if the modularity theorist were committed to the claim that children are innatelyprogrammed to develop one specific theory in a given cognitive domain, in much the same way that theyare innately programmed to develop a specific number of fingers or a specific configuration of major

blood vessels. But, of course, modularity theorists need make no such claim. Chomsky's account oflanguage acquisition is a paradigm case of what G&M would call a "modularity theory;" indeed, it is anexample that they frequently cite. Since there are estimated to be between 4,000 and 6,000 extantlanguages (Pinker, 1994, p. 232) and since there are obviously many more humanly possible languagesthat are no longer spoken or never have been, it is clearly quite wrong to suggest that modularity theoriesrequire that all sane and healthy children end up with the same theory regardless of the environment inwhich they are raised. G&M clearly recognize the problem, though they sometimes suggest that it's reallyjust a minor difficulty since the number of theories a modular account could produce must be rathersmall.

In some [modular] theories several alternative branching routes, so to speak, determine theeventual form the module may take. These are generally described as "parameters" set by theinput (Chomsky, 1986). Parameters allow for a somewhat richer developmental story thanthe one in which a module is simply turned on or off. The relation between the input and thesetting of the parameter is still, however, a relation of triggering. In contrast, in a theorytheory, by analogy with scientific theories, there should be indefinite scope for genuinelynovel theories, not simply a choice of several options. (G&M, p. 55)

But a bit of elementary math shows how odd it is to think that parameter setting models only offer "achoice of several options." Suppose that in language, or in some other cognitive domain, there are 15parameters that need to be set, and that each of these can take one of three values. On these veryconservative assumptions, the module would make 315 (or 14,348,907) options available. If we assume,not at all implausibly, that there are 25 parameters to set, the number of options rises to 847,288,609,443!

Once it is recognized that on modular accounts a child may have a huge number of options available, amuch deeper problem with G&M's "in principle" experiment comes into focus. There are imaginableoutcomes of the experiment that might indeed refute the theory theory, but there are no outcomes thatwould refute the modularity theory. So while the experiment might conceivably show that G&M's theoryis worse than the competition, it couldn't possibly show that their theory is better. To see the point, let usimagine that we raise a child on Vulcan, a planet with an extremely bizarre Star Trek-inspiredenvironment in which objects move, come into existence and go out of existence in ways that areradically different from the ways in which objects behave here on Earth. The experiment has two possibleoutcomes: either the child ultimately acquires a theory that accurately describes the principles governingthe behavior of objects in this bizarre world, or he does not. Neither of these outcomes poses a problemfor the modularity theorist. If the child does acquire a correct theory, the modularity theorist can explain itin much the same way that she explains the child's acquisition of a language that is radically differentfrom the language spoken by his biological parents. The language acquisition module makes millions ofoptions available, and if the language spoken on Vulcan happens to be one of these, then the child willacquire Vulcan just as smoothly as children in France acquire French and deaf children (in the appropriateenvironment) acquire ASL. Similarly, a modularity theorist might maintain that the theory-of-appearance-module makes millions of theories available, and if an accurate theory about object movements on Vulcanhappens to be one of them, then the child will acquire it. But, of course, even if the language acquisitionmodule makes millions of options available to the child, there will also be endlessly many logicallypossible languages that the module does not make available. And if a child is exposed to one of these, hewill not learn it, though he may end up speaking some hitherto unspoken language, in much the same waythat a child exposed to a pidgin ends up creating and speaking a creole. Analogously, if an accurate theoryof object movements on Vulcan is not among the theories that the theory-of-appearance-module makesavailable, then the child will not acquire it. He may either acquire some other theory (the analog ofacquiring a creole) or none at all. But whatever the outcome may be it will pose no problem for themodularity theorist, since her theory is compatible with all of the possible outcomes and does not predictany of them.

The situation is quite different for the theory theorist. If the child succeeds in acquiring an accurate theoryof object movements on Vulcan, all is well. For the child, according to the theory theorist, is a littlescientist equipped with "particularly powerful and flexible cognitive devices, devices that are good atconstructing accurate representations of new and unexpected worlds." (Ga, p. 490) So if science canfigure out the principles governing the behavior of objects on Vulcan, the child should be able to do ittoo. But now suppose that the child fails to learn an accurate theory of Vulcan object movements. Whataccount can the theory theorist offer in this case? Actually, there are a pair of sub-cases to consider. OnG&M's account, the mechanisms underlying scientific reasoning in both the scientist and the child arepowerful and flexible, but they are not completely unconstrained. "The theory theory, after all, stillassumes that not all the logically possible theories compatible with the evidence will actually beconstructed. There are some possible theories that will be constructed by human beings, given a particularpattern of evidence, and some that will not." (G&M, pp.55-56) Thus some logically possible worlds willbe so "new and unexpected" that human science simply cannot construct accurate theories about them. IfVulcan turns out to be a world like this, then all is well for the theory theorist. The child can't figure outwhat's going on and neither can the scientist. But what about the other possible sub-case, the one in whichhuman scientists can discover how things work on Vulcan but human children can't? Here, it would seem,the theory theorist is in trouble. For if the scientist and the child are using the same cognitive devices inconstructing their theories, then it is puzzling that scientists can understand a world and children can't.

It is important to see that our argument, in the last two paragraphs, does not in any way depend on theclaim that there are possible worlds whose laws or principles can be discovered by human scientists butnot by human children. What we are arguing is that there are some outcomes of G&M's "in principlesimple experiment" that would be problematic for the theory theorist, while there are no outcomes thatwould be problematic for the modularity theorist. Thus even if the experiment could be run, it could notpossibly show that the theory theory is better than the modularity theory.

The third suggestion: evidence vs. triggering

The third strategy that G&M propose for showing that the theory theory is preferable to modularitytheories focuses on the distinction between developmental processes that are triggered by someenvironmental input and those that treat the environmental input as evidence and produce rational changesin internalized theories or other representations in response to this evidence. The distinction has played avery central role in debates between nativists and empiricists in cognitive science during the last severaldecades. Despite this, however, the distinction is hardly a model of clarity, and in many cases it is farfrom obvious whether a process counts as triggered or as rational and evidence driven. This is becausethose who invoke the distinction typically explain it by appeal to prototypical examples, and the further aprocess is from one of the standard prototypes, the harder it is to classify. Scientific reasoning is, ofcourse, the favorite prototype of a rational, evidence driven process. As a first pass at explaining thenotion of triggering, many authors appeal to examples like imprinting. As Konrad Lorenz famouslydemonstrated, young goslings behave as though the first middle sized animate object that they see afterhatching is their mother. Normally, of course, that animate object is their mother, thus it makes perfectlygood evolutionary sense for them to be designed in this way. But there is nothing that much resemblesevidence driven scientific reasoning going on in the process that leads a gosling to believe (or behave as ifit believed) that the object in question is Mom. Rather, to use Fodor's memorable term, it is a "brutecausal process" (Fodor, 1981, p. 273) which when triggered by Lorenz rather than Mama Goose will leadthe goslings to parade around after him and ignore their mother.

Now, as G&M correctly note, modularity theorists typically claim that the processes which lead to theavailability and development of module based knowledge structures are brute causal triggerings, notrational and evidence driven. Thus if there is evidence that the processes driving the development of thetheory of object appearances, the theory of action and the theory of kinds are rational and evidence driven

rather than triggered, this will be a good reason to think that the theory theory is correct in these domains,and that modularity theories are false. But, as G&M are aware, a rational, evidence driven process "maynot be profoundly different from the case of a module with a great many parameters differently triggeredby evidence" (G&M, p. 56) Indeed, "[t]here is an interesting conceptual and formal question aboutwhether a modular system with a sufficiently varied set of parameters and triggers would reduce to a[science-like] theorizing system, or vice versa." (G&M, p. 55) Thus it is not surprising that in the finalpage or two of the chapters dealing with object appearances, actions and kinds, G&M admit quitecandidly that they have no "direct evidence" to offer for the claim that the developmental processesthey've described are rational responses to evidence rather than merely being triggered.

The theory theory proposes that the motivation for these changes comes from the infant'sobservations of the behavior of objects. It is the result of evidence. Yet again, we have nodirect experimental support for this claim. (G&M, p. 185; for similar passages see p. 122 &p. 160).

In each case G&M suggest that there may be some indirect evidence, which shows that extensiveexposure to relevant evidence, or making the relevant sorts of evidence more salient by encoding themlinguistically, can accelerate the acquisition of the theory. But it is hard to see how any of this evidencecould count in favor of the theory theory and against a modular theory that posits "a great many differentparameters differently triggered by evidence." (G&M, p. 56)

This brings us to the end of our assessment of G&M's arguments for the claim that the theory theory issuperior to the competition because it does a better job at explaining the available evidence. Theconclusion for which we've been arguing is that G&M's arguments do not even come close to makingtheir case. The argument which appeals to the static properties of theories is, by their own admission,entirely irrelevant since an advocate of a modularity account might well posit a mechanism which, whentriggered, yields representations with all of those static properties. The three arguments based on thedynamic properties of theories fare no better. The fact that the child's representations of the world changeand improve over time is not in the least embarrassing for the modularity theorist, since evolutionaryprocesses have made parallel patterns of change a ubiquitous feature of development. Speculations aboutwhat children would learn in a radically different universe are doubly unconvincing. First, of course, thereis no data. Second, no possible outcome of the hypothetical experiment could show that the modularitytheory is wrong, though there are outcomes that would be hard for the theory theory to explain. Finally, asG&M admit, they have no direct evidence that the processes of cognitive development in early childhoodare rational and evidence driven, as the theory theory requires, rather than triggered as the modularitytheorist would maintain. So the best that can be said for G&M's strong version of the theory theory is thatit is (more or less) compatible with a broad range of evidence. But much the same can be said formodularity theories, and also for mixed accounts like the one favored by Carey and Spelke in whichmodular core knowledge and science-like theory revision mechanisms both play a major role in cognitivedevelopment. Far from knocking out the competition, G&M have not even landed a solid blow.

5. Retro Philosophy of ScienceAs we noted in 4.1, many of the claims that G&M make about scientific theories are, and are intended tobe, "largely uncontroversial, not to say bland;" (Ga, p. 496) They want their portrait of scientific theoriesto be "compatible with many different philosophical accounts." (Ga, p. 495) But there is one cluster ofquestions about the underdetermination of theories by data and about the role of convention and socialfactors in science on which G&M depart from this strategy and adopt what Gopnik describes as an"unashamedly and self-consciously retro" position in the philosophy of science. (Gb, p. 560) The positionthey adopt is quite an extreme one which forces them to reinterpret or reject much of what has been done

in the philosophy, history and sociology of science during the last half century. However, it is also aposition which is, near enough, forced on them by their extreme, theories-all-the-way-down version of thetheory theory.

Since the time of Duhem, in the early years of the 20th century, it has been widely recognized in thephilosophy of science that theories are underdetermined by their evidence in the sense that, from the pointof view of deductive logic, any finite body of evidence is logically compatible with an indefinitely largerange of theories. In the middle years of the century, Carnap and other logical empiricists devoted a greatdeal of effort to developing non-deductive logics which could narrow the underdetermination byassessing the degree to which a given hypothesis was supported by a body of evidence. On Carnap'saccount, however, the use of an inductive logic presupposes the choice of a language or a "linguisticframework" which imposes significant constraints on the ontology of the theories set out in thatframework. In Carnap's "rational reconstruction" of scientific inquiry the process proceeds in two distinctstages: the first is the choice of a linguistic framework, the second is the formulation and testing ofhypotheses and theories within that framework. The initial choice of a framework, and any subsequentdecision to reject one framework and replace it with another "cannot be judged as being either true orfalse because it is not an assertion. It can only be judged as more or less expedient, fruitful, conducive tothe aim for which the language is intended." (Carnap, 1956, p. 207) Starting in the early 1950's, Quinemounted an enormously influential attack on Carnap's distinction between those parts of a theory thatwere mandated by the linguistic framework, and thus true by convention, and those that were left open bythe framework and accepted or rejected on the basis of evidence. The distinction is tenable, Quine argued,only if we can draw a principled distinction between analytic and synthetic sentences, and there is noprincipled distinction to be found. So while Carnap takes "a pragmatic stand on the question of choosingbetween language forms [or] scientific frameworks," Quine "espouse[s] a more thorough pragmatism."(Quine, 1953, p. 46) "Carnap maintains that ontological questions, and likewise questions of logical ormathematical principle, are questions not of fact but of choosing a convenient conceptual scheme orframework for science; and with this I agree only if the same be conceded for every scientifichypothesis." (Quine, 1966, p. 134)

As Gopnik sees it, all of this was a colossal mistake.

The worm in the apple, the serpent in the garden of these accounts was the idea, which is inboth Quine and Carnap, that the choice of the new language was "conventional" or"pragmatic." The implication, later made quite explicit in Quine, was that the decision wasalso arbitrary and simply socially determined…. This idea set the stage for the laterskepticism of Quine and the social constructionists who were to follow him. (Ga, p. 503)

While there are many objections that might be raised about Gopnik's interpretation of Carnap and Quineas early advocates of the view that much of science is "arbitrary and simply socially determined," theseare really not to the point. For whether or not Carnap and Quine held such views, there can be no disputethat in subsequent years many authors influenced by Quine(9) and many others who probably never read aword of Quine, have argued that some aspects of theory change in science are indeed arbitrary or aredetermined by a process of social negotiation in which both political factors and the personalities ofindividual scientists play a role. There is an enormous literature of case studies aimed at showing how atone or another point in the history of science the way in which a theory evolved or was replaced by aradically different theory was decisively influenced by social, political and psychological factors, andalso, at times, by chance.(10)

It is clear that if they are to maintain their strong version of the theory theory, G&M must reject all ofthese studies. For on their version of the theory theory the processes subserving theory change in scienceare identical to the processes subserving theory change in children - including very young children. But

obviously the social and political processes which, according to these studies, play a substantive role indetermining the content of new scientific theories in the course of theory revision are not processes thatplay any role in the development of theories in prelinguistic children. Nor do the arbitrary choices thatputatively influence the development of scientific theories, since if there were such arbitrary choices indevelopment, then we would expect different children to end up with significantly different theories aboutobject appearances, actions and kinds rather than all ending up, as G&M claim, with essentially the sametheories. G&M can and do recognize that social factors play a role in facilitating science, by enabling thecollection and dissemination of new evidence, for example. But their account of science as a "spandrel, anepiphenomenon of childhood," (Ga, p. 490) can't allow these factors to play any significant role indetermining the content of a new scientific theory which succeeds in replacing an old one. On their view,the content of a new theory is entirely determined by the old theory, the evidence, and the innateprocesses of theory revision.

The theory theory proposes that there are powerful cognitive processes that revise existingtheories in response to evidence. If cognitive agents began with the same initial theory, triedto solve the same problems, and were presented with similar patterns of evidence over thesame period of time they should, precisely, converge on the same theories at about the sametime. (Ga, p. 494)(11)

While it is clear that G&M's theory requires them to reject much of what has gone on in the history andsociology of science for the last few decades, we're not much impressed by the reasons they offer fordoing so. As we read them, the principal reason they offer is that if claims about social factors orconventions or arbitrary choices playing a substantive role in theory change were correct, it would bedifficult to explain why science leads to the truth.

[S]ocial interaction, by itself, can't produce veridical theories or genuine theory change…(G&M, p. 71)

Assimilating all cognitive development to the model of socialization is … a dreadful mistake,allied to the dreadful mistake of postmodernism in general. The crucial fact about cognitivedevelopment, and cognition in general, is that it is veridical, it gives us a better understandingof the world outside ourselves. Purely social-constructivist views discount this fundamentallink between the mind and the world. (G&M, p. 72)

The reference to "purely social-constructivist views" is, we think, a bit of rhetorical over kill. For while itis no doubt true that a pure social-constructivist view (whatever exactly that might be) would have a hardtime explaining how science succeeds in producing veridical theories, it is no easier to see how sciencecan produce veridical theories if social or political factors play any significant role in determining howtheory change works in science. But this can hardly be taken to be a serious reason to reject theseaccounts of science, since as Gopnik herself acknowledges, it is "profoundly mysterious" how anyprocess of theory change "generates representations that match up to the outside world." (Ga, p. 502)Since Gopnik provides no very persuasive justification for her "self-consciously retro" taste in thephilosophy of science, we think the link between those retro views and G&M's strong version of thetheory theory constitutes a real liability for their theory. Perhaps all those studies putatively showing thatsocial and political factors have played major roles in determining how scientific theories change are, likepostmodernism, just a dreadful mistake. But it will be no easy task to show that they are. And if even amodest number of those studies turn out to be correct, then G&M's version of the theory theory is simplymistaken.

References

Bloor, D. 1976. Knowledge and Social Imagery. Chicago: University of Chicago Press.

Carey, S. and Spelke, E. 1996. Science and Core Knowledge. Philosophy of Science, 63, 515-533.

Carnap, R. 1956. Empiricism, Semantics and Ontology. In Carnap, R., Meaning and Necessity. Chicago:University of Chicago Press. Pp. 205-221.

Carruthers, P. and Smith, P. 1996. Theories of Theories of Mind. Cambridge: Cambridge UniversityPress.

Chomsky, N. 1986. Knowledge of Language: Its Nature, Origins, and Use. New York: Praeger.

Davies, M. and Stone, T. (eds.), 1995a. Folk Psychology: The Theory of Mind Debate. Oxford:Blackwell.

Davies, M. and Stone, T. (eds.), 1995b. Mental Simulation: Evaluations and Applications. Oxford:Blackwell.

Deacon, T. 1997. The Symbolic Species. New York: W. W. Norton.

Downes, S. forthcoming. Constructivism. In The Routledge Encyclopedia of Philosophy. London:Routledge.

Feyerabend, P. 1981. Realism, Rationalism and Scientific Method. Cambridge: Cambridge UniversityPress.

Fodor, J. 1981. The Present Status of the Innateness Controversy. In Fodor, J., Representations.Cambridge, MA: MIT Press.

Galison, P. & Stump, D. 1996. The Disunity of Science. Stanford: Stanford University Press.

Giere, R. 1996. The Scientist as Adult. Philosophy of Science, 63, 538-541.

Godfrey-Smith, P. 1991. Teleonomy and the Philosophy of Mind. Ph.D. Dissertation. University ofCalifornia, San Diego.

Gopnik, A. 1984. Conceptual and Semantic Change in Scientists and Children: Why There Are NoSemantic Universals. Linguistics, 20, 163-179.

Gopnik, A. 1988. Conceptual and Semantic Development As Theory Change: The Case of ObjectPermanence. Mind and Language, 3, 197-216.

Gopnik, A. 1996a. The Scientist as Child. Philosophy of Science, 63, 485-514.

Gopnik, A. 1996b. A Reply to Commentators. Philosophy of Science, 63, 552-561.

Gopnik, A. and Meltzoff, A. 1997. Words, Thoughts and Theories. Cambridge, MA: MIT Press.

Gopnik, A. and Wellman, H. 1992. Why the Child's Theory of Mind Really Is a Theory. Mind andLanguage, 7, 145-171.

Gopnik, A. and Wellman, H. 1994. The Theory-Theory. In L. Hirschfeld and S. Gelman (eds.), Mappingthe Mind: Domain Specificity in Cognition and Culture. New York: Cambridge University Press. Pp.257-293.

Gould, S. 1977. Ontogeny and Phylogeny. Cambridge, MA: Harvard University Press.

Hinton, G. and Nolan, S. 1987. How Learning Can Guide Evolution. In Complex Systems, vol 1.Technical report CMU-CS-86-128. Carnegie-Mellon University. Pp. 495-502.

David Hull, 1990. Science as a Process. Chicago: University of Chicago Press.

Knorr-Cetina, K. and Mulkay, M. 1992. Science in Context. London: Sage.

Meltzoff, A. and Moore, M. 1983. Newborn Infants Imitate Adult Facial Gestures. Child Development,54, 702-709.

Meltzoff, A. and Moore, M. 1989. Imitation in Newborn Infants: Exploring the Range of GesturesImitated and the Underlying Mechanisms. Developmental Psychology, 25, 954-962.

Morton, A. 1980. Frames of Mind: Constraints on the Common-Sense Conception of the Mental. Oxford:Clarendon Press.

Pickering, A. 1984. Constructing Quarks. Chicago: University of Chicago Press.

Pinker, S. 1994. The Language Instinct. New York: William Morrow and Co.

Potts, R. 1996. Humanity's Descent: The Consequences of Ecological Instability. New York?: WilliamMorrow.

Quine, W. 1953. Two Dogmas of Empiricism. In Quine, W., From a Logical Point of View. Cambridge,MA: Harvard University Press. Pp. 20-46.

Quine, W. 1966. On Carnap's Views on Ontology. In Quine, W., The Ways of Paradox. New York:Random House. Pp. 126-134.

Ridley, Mark. 1993. Evolution. Cambridge, MA: Blackwell Science.

Shapin, S. and Shaffer, S. 1985. Leviathos and the Airpump. Princeton: Princeton University Press.

Spelke, E. 1994. Initial Knowledge: Six Suggestions. Cognition, 50, 431-445.

Spelke, E. Breinlinger, K., Macomber, J., and Jacobson, K. 1992. Origins of Knowledge, PsychologicalReview, 99, 605-632.

Stich, S. 1983. From Folk Psychology to Cognitive Science. Cambridge, MA: MIT Press.

Stich, S. 1996. Deconstructing the Mind. New York: Oxford University Press.

Stich, S. and Nichols, S. 1992. Folk Psychology: Simulation vs. Tacit Theory. Mind and Language, 7, 29-65.

Stich, S. and Nichols, S. 1995. Second Thoughts on Simulation. In Davies, M. and Stone, T. (eds.),Mental Simulation: Philosophical and Psychological Essays. Oxford: Blackwells. Pp. 87-108.

Wellman, H. 1985. The Child's Theory of Mind: The Development of Conceptions of Cognition. In S.Yussen (ed.), The Growth of Reflection in Children. Orlando: Academic Press.

Wellman, H. 1990. The Child's Theory of Mind. Cambridge, MA: MIT Press.

Wellman, H. and Gelman, S. Cognitive Development: Foundational Theories of Core Domains. AnnualReview of Psychology, 43, 337-375.

NOTES

1. We're grateful to Stephen Downes, Peter Godfrey-Smith and Jerry Fodor for a number of helpfulsuggestions.

2. Gopnik (1996a & 1996b). We will adopt the following abbreviations: Ga = Gopnik (1996a); Gb =Gopnik (1996b); G&M followed by a page number = Gopnik and Meltzoff (1997); G&M without pagenumbers = Gopnik and Meltzoff (the authors, not the book).

3. For more on the simulation theory, see the essays in Davies and Stone (1995a & 1995b) and inCarruthers and Smith (1996).

4. We should note that there is no substantive disagreement between G&M and ourselves here. In ourdefense of the theory theory in Stich & Nichols (1992 & 1995), we were concerned to argue that adultfolk psychological skills are subserved by a mentally represented knowledge structure -- a body of folkpsychological information -- rather than by a process of off-line simulation. And, of course, G&M agree.Given our polemical goals, there was no need for us to take a stand on how similar this mentallyrepresented knowledge structure is to a scientific theory, and thus we simply left the matter open.

5. The domains they mention include "folk economics" (G&M, p.174), "hunter-gatherer folk botany" and"Aborigine geography" (Gb, p. 561).

6. Indeed, there is reason to think that natural selection often proceeds in just the opposite direction. Theso-called "Baldwin effect" is a process in which Darwinian selection can mimic aspects of theLamarckian inheritance of acquired characteristics. In Baldwinian evolution, learned traits which arereliably adaptive in a given environment are gradually genetically assimilated because mutations thatresult in some or all of the previously learned information becoming innate will, under appropriatecircumstances, be favored by natural selection. For useful discussions of the Baldwin effect, see Godfrey-Smith (1991), Sec. 6.5, and Deacon (1997), Ch. 11. For an elegant and influential attempt to model theBaldwin effect, see Hinton & Nolan (1987).

7. Consider, for example, the following passage from Spelke et al. (1992) , p. 629: "These conceptions [ofmaterial objects] will be perpetuated over spontaneous development, because they serve to single out theobjects about which humans gain knowledge…. They can be overturned by instruction or disciplinedreflection if the student or scientist can use conceptions in a different domain of knowledge, such asmathematics, in order to single out a new set of entities in the physical world." (Emphasis added.)

8. The biogenetic law, that ontogeny always recapitulates phylogeny, has been widely rejected in biology(Gould 1977; Ridley 1993). There are numerous exceptions to the law. For instance, in some species,adults retain juvenile features of ancestral species (e.g., the Mexican axolotl, an aquatic salamander). Inother species, there are developmental stages that probably don't recapitulate any ancestral stage (Ridley1993, p. 539).

However, for our purposes the crucial point is simply that recapitulation is quite common. On this point,there is widespread agreement. For instance, after considering exceptions to the biogenetic law, Ridleywrites, "These exceptions not withstanding, recapitulation is noticeably common. Evolution has oftenproceeded by terminal addition" (Ridley 1993, p. 539).

9. Including one of the present authors, see Stich (1996), pp.63-82.

10. See, for example, Bloor (1976), Feyerabend (1981), Galison and Stump (1996), Hull (1990), Knorr-Cetina & Mulkay (1982), Pickering (1992), and Shapin & Shaffer (1985) For a helpful overview, seeDownes (forthcoming).

11. G&M do report data indicating developmental differences between Korean speaking and Englishspeaking children, and these differences seem to correspond in interesting ways to linguistic differencesbetween Korean and English. However, G&M do not argue that Korean and English speaking childrendevelop substantively different theories. Rather, the differences are restricted to the rate at which certainabilities develop (pp. 205-6).